The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.

The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.

An instruction to perform converting and scaling operations is provided. Execution of the instruction includes converting an input value in one format to provide a converted result in another format. The converted result is scaled to provide a scaled result. A result obtained from the scaled result is placed in a selected location. Further, an instruction to perform scaling and converting operations is provided. Execution of the instruction includes scaling an input value in one format to provide a scaled result and converting the scaled result from the one format to provide a converted result in another format. A result obtained from the converted result is placed in a selected location.

Data format conversion processing of an accelerator accessed by a processor of a computing environment is reduced. The processor and accelerator use different data formats, and the accelerator is configured to perform an input conversion to convert data from a processor data format to an accelerator data format prior to performing an operation using the data, and an output conversion to convert resultant data from accelerator data format back to processor data format after performing the operation. The reducing includes determining that adjoining operations of a process to run on the processor and accelerator are to be performed by the accelerator, where the adjoining operations include a source operation and destination operation. Further, the reducing includes identifying for removal output data format conversion of output data of the source operation for input to the destination operation, and input data format conversion of the input to the destination operation.

Embodiments are directed to systems and methods for reuse of FMA execution unit hardware logic to provide native support for execution of get exponent, get mantissa, and/or scale instructions within a GPU. These new instructions may be used to implement branch-free emulation algorithms for mathematical functions and analytic functions (e.g., transcendental functions) by detecting and handling various special case inputs within a pre-processing stage of the FMA execution unit, which allows the main dataflow of the FMA execution unit to be bypassed for such special cases. Since special cases are handled by the FMA execution unit, library functions emulating various functions, including, but not limited to logarithm, exponential, and division operations may be implemented with significantly fewer lines of machine-level code, thereby providing improved performance for HPC applications.

The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.

Methods and apparatus are disclosed supporting a design flow for developing quantized neural networks. In one example of the disclosed technology, a method includes quantizing a normal-precision floating-point neural network model into a quantized format. For example, the quantized format can be a block floating-point format, where two or more elements of tensors in the neural network share a common exponent. A set of test input is applied to a normal-precision flooding point model and the corresponding quantized model and the respective output tensors are compared. Based on this comparison, hyperparameters or other attributes of the neural networks can be adjusted. Further, quantization parameters determining the widths of data and selection of shared exponents for the block floating-point format can be selected. An adjusted, quantized neural network is retrained and programmed into a hardware accelerator.

An apparatus and method to execute ray tracing instructions. For example, one embodiment of an apparatus comprises execution circuitry to execute a dequantize instruction to convert a plurality of quantized data values to a plurality of dequantized data values, the dequantize instruction including a first source operand to identify a plurality of packed quantized data values in a source register and a destination operand to identify a destination register in which to store a plurality of packed dequantized data values, wherein the execution circuitry is to convert each packed quantized data value in the source register to a floating point value, to multiply the floating point value by a first value to generate a first product and to add the first product to a second value to generate a dequantized data value, and to store the dequantized data value in a packed data element location in the destination register.

Techniques for data type conversion using an instruction are described. An exemplary instruction includes fields for an opcode, an identification of source operands, and an identification of destination operand, wherein the opcode is to indicate execution circuitry and/or memory access circuitry is to convert 32-bit floating point values from the identified source operands into 16-bit floating point values and store 16-bit floating point values in data element positions of the identified destination operand.

Techniques for data type conversion are described. An example uses an instruction that is to include fields for an opcode, an identification of source operand location, and an identification of destination operand location, wherein the opcode is to indicate instruction processing circuitry is to convert a 16-bit floating-point value from the identified source operand location into a 32-bit floating point value and store that 32-bit floating point value in one or more data element positions of the identified destination operand.